Anthropic is adding its invisible text watermark to Claude models released before August 2, 2026. Mythos 5, Opus 4.7 and Sonnet 4.6 started on November 3, and Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 follow on November 16, according to The Decoder. Anthropic's help center lists no opt-out, and only regulators, media, researchers and obligated companies can check text for the mark.
If you write with Claude, assume the text now carries an invisible mark, whichever Claude model you pick. Anthropic is adding its text watermark to models it released before August 2, 2026, and its help center article offers no setting to switch it off. Nothing changes in how Claude’s answers read, and you don’t need to do anything to keep using it. What changes is who can later tell that Claude touched a piece of writing.
What Anthropic confirms
Anthropic’s help center says models launched on or after August 2, 2026 carried the watermark from launch, and that it “is adding watermarks to outputs from models released before August 2, 2026,” in line with the EU AI Act Code of Practice it signed. Its model table now lists text watermarks for Mythos 5, Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 4.6, Sonnet 4.5 and Haiku 4.5, alongside the current Fable, Mythos, Opus, Sonnet and Haiku models.
The mark is applied at the model level, so Anthropic says it is present “no matter which Claude product or surface the text comes from.” That covers the API, the Claude apps, Claude Code, Claude Cowork and Claude Tag, worldwide. Text from newer models carries the mark through AWS, Google Cloud and Microsoft Foundry too. For Mythos 5, Opus 4.7 and Sonnet 4.6, Anthropic says the Amazon Bedrock rollout “will complete by November 10.” Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 are listed with the mark on Anthropic’s own services only.
Anthropic says the mark is woven into the text, survives copy and paste, and “may persist through some editing.” Heavy editing, paraphrasing, translation, mixing with other writing or a very short passage can leave no detectable signal.
The dates come from Anthropic’s email
Anthropic’s help article gives no rollout dates. The Decoder reports them from an Anthropic email to API customers:
- November 3: Mythos 5, Opus 4.7 and Sonnet 4.6. Opus 4.6 was in this group but slipped, with no reason given.
- November 16: Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5.
The Decoder says customers can’t turn the watermark off, and quotes the email as saying the mark “contains no information about the user, their organization, or their conversations with Claude.” GIGAZINE reported the same November 3 start. We could not read the email itself.
Who can check text for the mark
You can’t. Text detection is in private preview and open only to groups the EU law names, such as regulators, law enforcement, media, fact-checkers, independent researchers and educational organizations, plus companies that must verify watermarks for their own compliance. Anthropic says it plans to widen access over time. Its free Content Checker reads Content Credentials on files Claude creates, such as PNG and JPEG images. It does not check text.
Anthropic is careful about what a hit means. A detected mark only says the text “may have been processed by Claude,” which includes proofreading, translating or summarizing your own writing. A missing mark proves nothing either way.
Developers can still lose the mark with one setting
On Sonnet 4.6, Anthropic’s help center says outputs “may not reliably carry the text watermark” when temperature, top_p or top_k is set to anything but the default, and it lists the same limit for Opus 4.6. The Decoder reports that Anthropic’s email gives the same condition for all four November 16 models and tells customers who rely on the mark for compliance to keep those settings at their defaults. If your product must show that its Claude text is marked, leave sampling settings alone on those models, or move to a newer model.
Ghostwritten client work now leaves a trace
For most people who use Claude at work, this changes nothing day to day. The mark does not identify you, and only approved organizations can look for it. The group it touches is anyone who hands Claude’s text to a client, an editor or a school as their own: agencies, freelancers, students and anyone drafting press releases or reports. A newsroom, a university or an EU regulator in the detection preview can now flag that text as Claude-processed, even when it came from a model released before the rule took effect. If AI help needs disclosing under your contract or your school’s policy, disclose it. Assuming an older model leaves no trace no longer holds after November 16.
For other recent Claude changes, see Claude Ultra’s removal of the five-hour limit and Claude Duet, Anthropic’s proactive agent.